Moldable Job Allocation for Handling Resource Fragmentation in Computational Grid - Cloud, Grid and High Performance Computing: Emerging Applications - page 70

Figure 6. Average performance over 120 different speed configurations

• normal distribution. adaptive processor al-

location with normal distribution of run-
time-estimation errors.

are stable and can significantly improve system
performance. For the details, among all the 120
cases, the proposed moldable allocation policies
with runtime estimation outperform the original
moldable policy in 108 cases. For the last two cases in Figure 5, we present

their worst-case data within the estimation-error
range from 10% to 100% with the step of 10%.
The results in Figure 5 show that Grid computing
with moldable job allocation can greatly improve
the system performance compared to the non-Grid
architecture. Moreover, the improved moldable
job allocation policies with runtime estimation
can improve the system performance further
compared to the original moldable job allocation
policy. The results also indicate that estimation
errors lead to little influence on overall system
performance. Therefore, the proposed moldable
allocation policies are stable in a heterogeneous
computational Grid.
Figure 5 represents only one possible speed
configuration in a heterogeneous computational
Grid environment. To further investigate the ef-
fectiveness of the proposed policies, we conducted
a series of 120-case simulations corresponding to
all possible permutations of the site speed vec-
tor (1,3,5,7,9) under the SDSC's SP2 workload.
Figure 6 shows the average turnaround times over
the 120 cases for the five allocation policies in
Figure 5, accordingly. The results again confirm
that the proposed moldable job allocation policies

COMPARISON WITH MULTI-SITE
CO-ALLOCATION

Multi-site co-allocation (Sonmez, Mohamed, &
Epema, 2010) is another approach usually used
to deal with the resource fragmentation issue in
computational Grid environments. It allows a
parallel job to run across site boundary, simultane-
ously using processors from more than one sites.
Figure 7 compares multi-site co-allocation and
moldable job allocation under the SDSC's SP2
workload. In our job model, each job is associ-
ated with an attribute, slowdown , which indicates
how long its runtime would be extended to when
running with multi-site co-allocation in the Grid.
In the simulations, the slowdown values for these
jobs are generated according to specified statis-
tical distributions and upper limits. The upper
limits are denoted by p in Figure 5. Two types
of statistical distributions, uniform and normal
distributions, are evaluated in the simulations.
Results in Figure 5 show that the performance of
multi-site co-allocation is greatly affected by the